Why can't I get the attribute 'pos' data from a new model trained from scratch?

[qingyun1988]

This is my training data example which contains attribute 'pos' data and uses for training a blank model:

Example.from_dict(_nlp.make_doc('I love you.'), {'heads': [1, 1, 1, 1], 'deps': ['nsubj', 'ROOT', 'dobj', 'punct'],'words': ['I', 'love', 'you', '.'],'pos': ['PRON', 'VERB', 'PRON', 'PUNCT'],'tags': ['PRP', 'VBP', 'PRP', '.']})

But I can't get the attribute 'pos' data from a new model trained from scratch, which is empty:

Why? Is the 'pos' attribute data format in the data example wrong?

Is there anyone who would like to answer my question? Thank you!

[adrianeboyd]

Hi, does your model pipeline include a morphologizer? The tagger only annotates token.tag, not token.pos. The morphologizer annotates token.pos and token.morph.

If this isn't the case, could you post the full config and details about which version of spacy you're using?

[qingyun1988]

Thank you very much! I've solved it through your tips above. The version of Spacy I'm using is v3.0 and the training way I am using is API, not config.

I am very clear about how to train token.tag and token.pos on a model after you reminded me. Thank you again.

By the way, I want to consult you a question again.

This question is about resume-training a pre-trained model.

I used 100 examples that are all "so...that..." inversion sentences to resume training on the pre-trained model 'en_core_web_sm-3.0', but there is a little change in terms of dependencies which is not what I expect.

For example, there is a sentence:
So unreasonable was his price that everybody startled.

The view before resume-training:

The view I expect:

i.e. I expect the head of the word 'startled' is 'unreasonable', not 'was'.

But, the view after the model has trained:

The head of the word 'startled' didn't change at all, and the dep of the word 'startled' was replaced 'det' from the previous 'ccomp' .

Why is this change? Is the number of examples used for training not enough? Or did I do something wrong?

I expect your reply for me, thank you.

[svlandeg]

It looks like your follow-up question is the same as the one from #8295. It's probably better to keep things in one place.

[qingyun1988]

It looks like your follow-up question is the same as the one from #8295. It's probably better to keep things in one place.

@svlandeg There is a little difference between this question and the former one #8295. In #8295, the model I've trained on is 'lg' and in this question, the model is 'sm' which is far smaller than the 'lg' one. So, the examples I prepared should theoretically be enough for having some change, but the consequence is not the same as I expected.

My question is, were these examples not enough or did I do something wrong? Thank you!